Add tensor parallel support to T5 via NxD

benchmark/text-generation-inference/performance/generate_csv.py

@@ -3,7 +3,6 @@
 import os
 
 import pandas as pd
-
 from guidellm.core import GuidanceReport, TextGenerationBenchmark
 
 
@@ -16,11 +15,7 @@ def _benchmark_rate_id(benchmark: TextGenerationBenchmark) -> str:
     :return: A string representing the benchmark rate ID.
     :rtype: str
     """
-    rate_id = (
-        f"{benchmark.mode}@{benchmark.rate:.2f} req/sec"
-        if benchmark.rate
-        else f"{benchmark.mode}"
-    )
+    rate_id = f"{benchmark.mode}@{benchmark.rate:.2f} req/sec" if benchmark.rate else f"{benchmark.mode}"
     return rate_id
 
 
@@ -38,20 +33,20 @@ def main():
     for path in paths:
         filename = os.path.basename(path)
         # Extract model_id
-        model_id, date = filename.replace(suffix, '').split('#')
+        model_id, date = filename.replace(suffix, "").split("#")
         with open(path) as f:
             report = GuidanceReport.from_json(f.read())
             for benchmark in report.benchmarks:
                 for b in benchmark.benchmarks_sorted:
                     d = {
-                    "model_id": model_id,
-                    "Date": date,
-                    "Input type": _benchmark_rate_id(b),
-                    "Requests per Second": b.completed_request_rate,
-                    "Request Latency (s)": b.request_latency,
-                    "Time-to-first-token (ms)": b.time_to_first_token,
-                    "Inter Token Latency (ms)": b.inter_token_latency,
-                    "Output Token Throughput (t/s)": b.output_token_throughput,
+                        "model_id": model_id,
+                        "Date": date,
+                        "Input type": _benchmark_rate_id(b),
+                        "Requests per Second": b.completed_request_rate,
+                        "Request Latency (s)": b.request_latency,
+                        "Time-to-first-token (ms)": b.time_to_first_token,
+                        "Inter Token Latency (ms)": b.inter_token_latency,
+                        "Output Token Throughput (t/s)": b.output_token_throughput,
                     }
                     results.append(pd.DataFrame.from_dict(d, orient="index").transpose())
 

docs/source/package_reference/modeling.mdx

@@ -68,6 +68,11 @@ The following Neuron model classes are available for natural language processing
 [[autodoc]] modeling.NeuronModelForCausalLM
     - forward
 
+### NeuronModelForSeq2SeqLM
+
+[[autodoc]] modeling_seq2seq.NeuronModelForSeq2SeqLM
+    - forward
+
 ## Computer Vision
 
 The following Neuron model classes are available for computer vision tasks.

optimum/commands/export/neuronx.py

@@ -112,6 +112,12 @@ def parse_args_neuronx(parser: "ArgumentParser"):
         choices=["bf16", "fp16", "tf32"],
         help='The data type to cast FP32 operations to when auto-cast mode is enabled. Can be `"bf16"`, `"fp16"` or `"tf32"`.',
     )
+    optional_group.add_argument(
+        "--tensor_parallel_size",
+        type=int,
+        default=1,
+        help="Tensor parallelism size, the number of neuron cores on which to shard the model.",
+    )
     optional_group.add_argument(
         "--dynamic-batch-size",
         action="store_true",

optimum/exporters/neuron/__main__.py

@@ -264,6 +264,7 @@ def get_submodels_and_neuron_configs(
     task: str,
     output: Path,
     library_name: str,
+    tensor_parallel_size: int = 1,
     subfolder: str = "",
     dynamic_batch_size: bool = False,
     model_name_or_path: Optional[Union[str, Path]] = None,
@@ -300,7 +301,14 @@ def get_submodels_and_neuron_configs(
     elif is_encoder_decoder:
         optional_outputs = {"output_attentions": output_attentions, "output_hidden_states": output_hidden_states}
         models_and_neuron_configs, output_model_names = _get_submodels_and_neuron_configs_for_encoder_decoder(
-            model, input_shapes, task, output, dynamic_batch_size, model_name_or_path, **optional_outputs
+            model=model,
+            input_shapes=input_shapes,
+            tensor_parallel_size=tensor_parallel_size,
+            task=task,
+            output=output,
+            dynamic_batch_size=dynamic_batch_size,
+            model_name_or_path=model_name_or_path,
+            **optional_outputs,
         )
     else:
         # TODO: Enable optional outputs for encoders
@@ -427,6 +435,7 @@ def _get_submodels_and_neuron_configs_for_stable_diffusion(
 def _get_submodels_and_neuron_configs_for_encoder_decoder(
     model: "PreTrainedModel",
     input_shapes: Dict[str, int],
+    tensor_parallel_size: int,
     task: str,
     output: Path,
     dynamic_batch_size: bool = False,
@@ -442,15 +451,19 @@ def _get_submodels_and_neuron_configs_for_encoder_decoder(
     models_and_neuron_configs = get_encoder_decoder_models_for_export(
         model=model,
         task=task,
+        tensor_parallel_size=tensor_parallel_size,
         dynamic_batch_size=dynamic_batch_size,
         input_shapes=input_shapes,
         output_attentions=output_attentions,
         output_hidden_states=output_hidden_states,
+        model_name_or_path=model_name_or_path,
     )
     output_model_names = {
         ENCODER_NAME: os.path.join(ENCODER_NAME, NEURON_FILE_NAME),
         DECODER_NAME: os.path.join(DECODER_NAME, NEURON_FILE_NAME),
     }
+    model.config.save_pretrained(output)
+    model.generation_config.save_pretrained(output)
     maybe_save_preprocessors(model_name_or_path, output)
 
     return models_and_neuron_configs, output_model_names
@@ -475,6 +488,7 @@ def load_models_and_neuron_configs(
     lora_weight_names: Optional[Union[str, List[str]]],
     lora_adapter_names: Optional[Union[str, List[str]]],
     lora_scales: Optional[Union[float, List[float]]],
+    tensor_parallel_size: int = 1,
     controlnet_ids: Optional[Union[str, List[str]]] = None,
     output_attentions: bool = False,
     output_hidden_states: bool = False,
@@ -499,6 +513,7 @@ def load_models_and_neuron_configs(
     models_and_neuron_configs, output_model_names = get_submodels_and_neuron_configs(
         model=model,
         input_shapes=input_shapes,
+        tensor_parallel_size=tensor_parallel_size,
         task=task,
         library_name=library_name,
         output=output,
@@ -522,6 +537,7 @@ def main_export(
     model_name_or_path: str,
     output: Union[str, Path],
     compiler_kwargs: Dict[str, Any],
+    tensor_parallel_size: int = 1,
     model: Optional[Union["PreTrainedModel", "ModelMixin"]] = None,
     task: str = "auto",
     dynamic_batch_size: bool = False,
@@ -563,6 +579,7 @@ def main_export(
         model_name_or_path=model_name_or_path,
         output=output,
         model=model,
+        tensor_parallel_size=tensor_parallel_size,
         task=task,
         dynamic_batch_size=dynamic_batch_size,
         cache_dir=cache_dir,
@@ -597,6 +614,12 @@ def main_export(
     )
 
     # Validate compiled model
+    if do_validation and tensor_parallel_size > 1:
+        # TODO: support the validation of tp models.
+        logger.warning(
+            "The validation is not yet supported for tensor parallel model, the validation will be turned off."
+        )
+        do_validation = False
     if do_validation is True:
         try:
             validate_models_outputs(
@@ -698,6 +721,7 @@ def main():
         model_name_or_path=args.model,
         output=args.output,
         compiler_kwargs=compiler_kwargs,
+        tensor_parallel_size=args.tensor_parallel_size,
         task=task,
         dynamic_batch_size=args.dynamic_batch_size,
         atol=args.atol,

optimum/exporters/neuron/base.py

@@ -146,6 +146,7 @@ def __init__(
         task: str,
         compiler_type: Optional[str] = None,
         compiler_version: Optional[str] = None,
+        tensor_parallel_size: int = 1,
         batch_size: Optional[int] = None,
         text_batch_size: Optional[int] = None,
         image_batch_size: Optional[int] = None,
@@ -174,6 +175,7 @@ def __init__(
         self._config = config
         self._normalized_config = self.NORMALIZED_CONFIG_CLASS(self._config)
         self.mandatory_axes = ()
+        self.tensor_parallel_size = tensor_parallel_size
         self.task = task
         self._axes: Dict[str, int] = {}
         self.dynamic_batch_size = dynamic_batch_size
@@ -227,6 +229,14 @@ def task(self, value: str):
         self._task = value
         self.mandatory_axes = self.get_mandatory_axes_for_task(self.task)
 
+    @property
+    def tensor_parallel_size(self) -> int:
+        return self._tensor_parallel_size
+
+    @tensor_parallel_size.setter
+    def tensor_parallel_size(self, value: int):
+        self._tensor_parallel_size = value
+
     def __getattr__(self, attr_name) -> Any:
         if attr_name != "_axes" and attr_name in self._axes:
             return self._axes[attr_name]